Why Units Matter in a Regression Equation
A regression coefficient is a number with a meaning tied to the units of the variables in the model. A slope of 4.5 does not, by itself, say how much an outcome changes: the interpretation depends on whether the outcome is measured in minutes or hours and whether the predictor is measured in kilometres or miles. Writing units explicitly helps keep the equation, predictions, and interpretations consistent.
Dimensional analysis means checking the measurement units of quantities and equations. In a regression equation, it provides a basic consistency test: terms that are added must have the same units, and a slope must convert a change in predictor units into a change in outcome units. This check does not establish that a model is appropriate or that an association is causal. It checks whether the units in the model make sense.
Units in an Additive Regression Model
Consider a fitted additive regression model with an outcome \(Y\) and \(p\) predictors:
Here, \(\hat{y}\) is the fitted outcome for a specified set of predictor values, \(b_0\) is the intercept, \(b_j\) is the coefficient of predictor \(x_j\), and \(p\) is the number of predictors. The equation describes the fitted value, not necessarily an observed outcome. If an observed outcome is written as \(y\), its residual is \(y-\hat{y}\); both \(y\) and the residual have the units of the outcome.
For example, if the outcome is measured in minutes and a predictor is measured in kilometres, the slope has units of minutes per kilometre. Multiplying that slope by a distance in kilometres gives a fitted contribution in minutes. The intercept also has units of minutes: an intercept is a term in an equation for a fitted outcome, not a unit-free constant.
A predictor may be a count, a continuous measurement, or an indicator coded 0 and 1. A binary indicator has no physical measurement unit: its values are dimensionless codes. Its coefficient therefore has the outcome units. The coefficient describes the fitted difference associated with changing the indicator from 0 to 1 while holding the other modeled predictors fixed.
The model’s structure and measurement units are enough to determine these unit relationships. Dimensional analysis does not require assumptions about the distribution of errors or a hypothesis test. It also does not check other modeling conditions, such as whether a linear form is suitable for the relationship being modeled.
A Short Derivation of Slope Units
The unit rule for a slope follows directly from the meaning of a slope as fitted change in the outcome per change in the predictor. In a model that holds all other predictors fixed, changing \(x_j\) by \(\Delta x_j\) changes the fitted value by \(b_j\Delta x_j\). The fitted change must be measured in outcome units. Dividing those units by the units of \(\Delta x_j\) gives the units of \(b_j\):
This also checks the whole equation. Each product \(b_jx_j\) has outcome units, and the intercept has outcome units, so every term can be added to produce a fitted value in outcome units. If one term has different units, either the stated coefficient units, the predictor units, or the equation itself has been misidentified.
Worked Example: Travel Time Per Kilometre
Suppose a hypothetical fitted model predicts a walking route’s travel time \(Y\), in minutes, from its length \(x\), in kilometres:
Step 1: Identify the outcome and predictor units. The fitted outcome is travel time in minutes, and \(x\) is route length in kilometres.
Step 2: Assign units to the coefficients. The intercept, 12.0, has units of minutes. The slope, 4.5, has units of minutes per kilometre, because multiplying it by \(x\) in kilometres must give a time in minutes.
Step 3: Check a prediction. For a hypothetical route of 6 kilometres:
The units cancel in the product because kilometres in the slope’s denominator cancel the kilometres in the route length. Both addends are minutes, so the prediction is 39.0 minutes. The slope’s interpretation is a fitted increase of 4.5 minutes per additional kilometre, not simply “4.5 minutes” without specifying the predictor change.
Checking Each Term in a Multiple Regression
With multiple predictors, each slope can have a different compound unit. The units are found separately for each predictor by dividing the outcome units by that predictor’s units. The fact that the numerical coefficients appear in one equation does not mean their units are alike. What must match are the units of the terms after each coefficient is multiplied by its predictor.
Worked Example: Weekly Cost From Area and Occupancy
Consider a hypothetical model for a building’s weekly heating cost \(Y\), measured in dollars per week. It uses floor area \(A\), measured in square metres, and \(N\), the number of occupants:
Step 1: Determine coefficient units. The intercept is 18 dollars per week. Since \(A\) is measured in square metres, \(0.80\) has units of dollars per week per square metre. Since \(N\) counts occupants, \(6.5\) has units of dollars per week per occupant. The count is a number of people; stating “per occupant” makes the change being described explicit.
Step 2: Check the equation’s terms. The product \(0.80A\) has units of dollars per week, as does \(6.5N\). Both can be added to the intercept, which is also in dollars per week.
Step 3: Calculate and check a fitted value. For a hypothetical building with area \(70\ \text{m}^2\) and 3 occupants:
Each predictor term contributes an amount in dollars per week, so the fitted value has those units. In this additive model, the area slope describes a fitted change of 0.80 dollars per week for each additional square metre, holding the number of occupants fixed. The occupancy slope describes a fitted change of 6.5 dollars per week for each additional occupant, holding area fixed. Neither interpretation alone establishes a causal effect.
Indicators and Unitless Predictors
A 0–1 indicator is useful for representing group membership. Although its values are numerical, the codes themselves are dimensionless. Its coefficient therefore has the same units as the outcome. Dimensional consistency still applies: multiplying a coefficient measured in outcome units by a unitless indicator leaves a term measured in outcome units.
Worked Example: Checking an Indicator Coefficient
Suppose a hypothetical model predicts a quiz score \(Y\), measured in points, from study time \(h\), measured in hours, and an indicator \(I\) that equals 1 for a student attending a review session and 0 otherwise:
Step 1: Assign units. The intercept, 42, is in points. The study-time coefficient, 3.2, is in points per hour. The indicator is dimensionless, so its coefficient, 8.0, is in points.
Step 2: Interpret the indicator term. When \(I\) changes from 0 to 1, the fitted value changes by 8.0 points, with study time held fixed. The coefficient is not 8.0 points per hour, because the indicator does not measure time. The study-time slope, by contrast, is a fitted change of 3.2 points per additional hour, holding the indicator fixed.
Step 3: Check a prediction. For a hypothetical student studying 5 hours and attending the session, \(h=5\) and \(I=1\):
All three terms are in points. If the indicator had instead been coded as 0 for attending and 1 for not attending, the units would remain the same, although the group comparison described by the coefficient would be reversed.
Finding and Correcting a Unit Mismatch
Dimensional checks are particularly useful when reading a coefficient table or translating a formula into words. A common error is to repeat the outcome units after every coefficient, without considering the predictor. Another is to multiply a slope by a predictor value without checking that the predictor’s units cancel. The equation gives a systematic way to catch both mistakes.
For example, a response measured in litres gives fitted values and intercepts in litres.
Record each predictor’s measurement units, or note that an indicator is dimensionless.
For each slope, divide the outcome units by that predictor’s units.
Multiply each slope’s units by its predictor’s units. Every resulting term must match the outcome units before the terms are added.
For instance, if a volume outcome is measured in litres and a predictor is measured in minutes, the slope must be in litres per minute. Calling the slope “litres” loses the rate information: a value of 2 litres per minute describes a fitted change for each minute, not a fitted change of 2 litres for any unspecified predictor change.
This check concerns units, not numerical plausibility. A dimensionally consistent equation can still have poorly estimated coefficients, an unsuitable functional form, or predictions outside a useful range. Conversely, a plausible-looking prediction may have been produced by an equation with a unit mistake. Both the arithmetic and the model’s broader suitability require attention.
Common Mistakes and Why the Check Matters
- Calling every coefficient an outcome-unit quantity. The intercept has outcome units, but a slope generally has outcome units per predictor unit. Include the predictor unit in the slope interpretation.
- Adding unlike quantities. A term in minutes cannot be added to a term in hours without expressing them on compatible units. In a regression equation, each term must match the outcome units.
- Giving a unitless interpretation to a count predictor. A count is numerical, but its slope describes change per additional counted item. “Dollars per occupant” is more informative than “dollars per one unit.”
- Assigning a rate unit to an indicator coefficient. A 0–1 indicator has no physical unit, so its coefficient has outcome units and represents a fitted group difference for a one-unit change in the indicator.
- Treating a dimensional check as evidence of causation or model quality. Consistent units are necessary for a coherent equation, but they do not show that changing a predictor causes a change in the outcome or that the model assumptions hold.
Check Your Understanding
Use the outcome and predictor units to answer each question.
- A model predicts distance in kilometres from time in hours. What are the units of the fitted outcome, the intercept, and the slope?
- A model predicts monthly electricity cost in dollars from floor area in square metres and a 0–1 indicator for apartment type. What are the units of each slope?
- A coefficient is reported as 5.0 points, and its predictor is study time in hours. What unit should the coefficient have if the outcome is a score in points? What is wrong with the reported unit?
- Explain why a term involving a 0–1 indicator has outcome units even though the indicator is represented by a number.
- A model has an outcome in kilograms and a predictor in years. Describe the units of the slope and explain what the slope represents in context.